gilliaan - CenterWide
Phantom center extraction models trained by me for use with MSST. Each one splits a stereo mix into two stems: center (what sits in the phantom center. Typically lead vocals, bass, kick, snare) and wide (everything panned off-center). Unlike a plain mid/side matrix this is a learned separation, so hard-panned and partially panned content ends up in wide instead of leaking into both.
Same task, four architectures. Pick whichever sounds best to you. The v3 models are the most recent and sound the best.
Each folder contains the .ckpt and its matching .yaml config.
Models
| Model | Arch | Size |
|---|---|---|
MelbandRoformer/gilliaan_melroformer_centerwide_v3 |
Mel-Band RoFormer | 870 MB |
MelbandRoformer/gilliaan_melroformer_centerwide_v2 |
Mel-Band RoFormer | 870 MB |
MelbandRoformer/gilliaan_melroformer_centerwide_v1 |
Mel-Band RoFormer | 870 MB |
BandSplitRoformer/gilliaan_bsroformer_centerwide_v3beta |
BS-RoFormer | 303 MB |
BandSplitRoformer/gilliaan_bsroformer_centerwide_v2 |
BS-RoFormer | 204 MB |
BandSplitRoformer/gilliaan_bsroformer_centerwide_v1 |
BS-RoFormer | 204 MB |
MDX23C/gilliaan_mdx23c_centerwide_v3beta |
MDX23C | 448 MB |
MDX23C/gilliaan_mdx23c_centerwide_v2e |
MDX23C | 440 MB |
MDX23C/gilliaan_mdx23c_centerwide_v1 |
MDX23C | 438 MB |
SCNet/gilliaan_scnet_centerwide_v2 |
SCNet | 214 MB |
SCNet/gilliaan_scnet_centerwide_v1 |
SCNet | 214 MB |
All output center and wide. Sum them back together and you get the original mix.